Energy-Efficient VMs
Placement
Albert De LaFuente
VigliottiDaniel
MacedoBatista
The Problem
The objective
Motivation
Related Work
ThepyCloudSimFramework
Experiments
Results andConclusions
References
Energy-Efficient Virtual Machines Placement
Albert De La Fuente VigliottiDaniel Macedo Batista
Department of Computer ScienceUniversity of Sao Paulo
albert at ime.usp.br
http://www.ime.usp.br/~albert — http://www.albertdelafuente.com
May 6th, 2014
1 / 24
Energy-Efficient VMs
Placement
Albert De LaFuente
VigliottiDaniel
MacedoBatista
The Problem
The objective
Motivation
Related Work
ThepyCloudSimFramework
Experiments
Results andConclusions
References
The problem
The current IT infrastructure contributes about 2% oftotal world wide power consumption and CO2 footprints[1].
This corresponds to the typical yearly electricityconsumption of 120 million households [1].
An energy consumption rise of 16-20% per year can beobserved in the last years on data centers and large-scalecomputing infrastructures, corresponding to a doublingevery 4-5 years [2].
2 / 24
Energy-Efficient VMs
Placement
Albert De LaFuente
VigliottiDaniel
MacedoBatista
The Problem
The objective
Motivation
Related Work
ThepyCloudSimFramework
Experiments
Results andConclusions
References
The objective
Question:
Is it possible to reduce the amount of consumed energy in adata center by using virtualization?
Question:
Is there reduction of energy consumption when keeping a samenumber of virtual machines in a lower number of physicalmachines?
Our approaches:
A Knapsack based algorithmAn Evolutionary Computation (EC) based algorithm
3 / 24
Energy-Efficient VMs
Placement
Albert De LaFuente
VigliottiDaniel
MacedoBatista
The Problem
The objective
Motivation
Related Work
ThepyCloudSimFramework
Experiments
Results andConclusions
References
Motivation
4 / 24
Energy-Efficient VMs
Placement
Albert De LaFuente
VigliottiDaniel
MacedoBatista
The Problem
The objective
Motivation
Related Work
ThepyCloudSimFramework
Experiments
Results andConclusions
References
Related work
CV Xavier et al. [3] analyzed performance, but they focusedonly on high performance computing environments (HPC).
CV Mehnert et al. [4] focused on memory incrementalcheckpointing (related on the EU-funded projectXtreemOS).
HV Beloglazov et al. [5] created OpenStack Neat which is anopen source software framework for distributed dynamicVM consolidation in cloud data centers based on theOpenStack platform.
5 / 24
Energy-Efficient VMs
Placement
Albert De LaFuente
VigliottiDaniel
MacedoBatista
The Problem
The objective
Motivation
Related Work
ThepyCloudSimFramework
Experiments
Results andConclusions
References
Related work: CloudSim
HV Calheiros et al. [6] created a simulation toolkit calledCloudSim It abstracts the low level details related toCloud-based infrastructures and services, allowing to focuson specific system design. It supports modeling andsimulation of:
Large scale Cloud computing data centersVirtualized server hosts, with customizable policies forprovisioning host resources to virtual machinesEnergy-aware computational resourcesData center network topologies and message-passingapplicationsFederated clouds
6 / 24
Energy-Efficient VMs
Placement
Albert De LaFuente
VigliottiDaniel
MacedoBatista
The Problem
The objective
Motivation
Related Work
ThepyCloudSimFramework
Experiments
Results andConclusions
References
The core of the simulation framework ofpyCloudSim
The main algorithm of pyCloudSim 1 iterates over the available(unplaced) physical hosts and VMs to determine a placementusing a given strategy S.
1https://github.com/vonpupp/sbrc-2014-simulation
7 / 24
Energy-Efficient VMs
Placement
Albert De LaFuente
VigliottiDaniel
MacedoBatista
The Problem
The objective
Motivation
Related Work
ThepyCloudSimFramework
Experiments
Results andConclusions
References
The Knapsack (KSP) based strategy
A list of constraints is built for each resource, this includesassigning a weight on each VM which will be the criteria to bemaximized by the algorithm, equivalent to maximize thenumber of VMs.
8 / 24
Energy-Efficient VMs
Placement
Albert De LaFuente
VigliottiDaniel
MacedoBatista
The Problem
The objective
Motivation
Related Work
ThepyCloudSimFramework
Experiments
Results andConclusions
References
The Evolutionary Computation (EC) based strategy
G generates possible solutions with 1% of chance to include aVM in a host. The evaluation function E calculates the fittingof the proposed solution. We used a population size of 50, atournament size of 25 and 2500 evaluations.
9 / 24
Energy-Efficient VMs
Placement
Albert De LaFuente
VigliottiDaniel
MacedoBatista
The Problem
The objective
Motivation
Related Work
ThepyCloudSimFramework
Experiments
Results andConclusions
References
The evaluation function (a valid solution)
number of VMs = 4
[ 70 70 60 60 ]-100
[ -30 -30 -40 -40 ]max(0, [ -30 -30 -40 -40 ]
[ 0 0 0 0 ]sum([ 0 0 0 0 ]
0
4 - 0 = 4
10 / 24
Energy-Efficient VMs
Placement
Albert De LaFuente
VigliottiDaniel
MacedoBatista
The Problem
The objective
Motivation
Related Work
ThepyCloudSimFramework
Experiments
Results andConclusions
References
The evaluation function (an invalid solution)
number of VMs = 4
[ 130 70 60 60 ]-100
[ +30 -30 -40 -40 ]max(0, [ +30 -30 -40 -40 ]
[ +30 0 0 0 ]sum([ +30 0 0 0 ]
+30
4 - 30 = -26
11 / 24
Energy-Efficient VMs
Placement
Albert De LaFuente
VigliottiDaniel
MacedoBatista
The Problem
The objective
Motivation
Related Work
ThepyCloudSimFramework
Experiments
Results andConclusions
References
The trace analysis
We analyzed more than 11,776 real traces (24-hour long each)from the PlanetLab project. The Standard deviation range wasfrom 0.2634 to 43.5875, and the mean range was from 0.5173to 95.9756. These values represents percentage of use.
12 / 24
Energy-Efficient VMs
Placement
Albert De LaFuente
VigliottiDaniel
MacedoBatista
The Problem
The objective
Motivation
Related Work
ThepyCloudSimFramework
Experiments
Results andConclusions
References
The methodology
A simulation is made of:
trace-scenario
algorithm-scenario [Energy Unaware, Iterated-KSP,Iterated-EC]
physical machine-scenario [10, 100], increments by 10
VMs varying on the interval [16, 288], increments by 16
We repeated each simulation 30 times to check if there was aclear tendency, and later reduced the data to three cases:
best case
worst case
average case
The experiment took ∼180h (more than one week).
13 / 24
Energy-Efficient VMs
Placement
Albert De LaFuente
VigliottiDaniel
MacedoBatista
The Problem
The objective
Motivation
Related Work
ThepyCloudSimFramework
Experiments
Results andConclusions
References
Power consumption comparison - Trace 1 (15.5204,25.0694)
Figure : Power consumption [100 hosts / Trace 1]
14 / 24
Energy-Efficient VMs
Placement
Albert De LaFuente
VigliottiDaniel
MacedoBatista
The Problem
The objective
Motivation
Related Work
ThepyCloudSimFramework
Experiments
Results andConclusions
References
Power consumption comparison - Trace 1 (15.5204,25.0694)
Figure : Power consumption [200 hosts / Trace 1]
15 / 24
Energy-Efficient VMs
Placement
Albert De LaFuente
VigliottiDaniel
MacedoBatista
The Problem
The objective
Motivation
Related Work
ThepyCloudSimFramework
Experiments
Results andConclusions
References
Conclusions - Power consumption
Iterated-EC had power savings starting from 35.46% for aworkload of 288 VMs and up to 92.20% for a workload of16 VMs with 200 hosts.
The Iterated-KSP had power savings starting from40.33% for a workload of 288 VMs and up to 92.21%for a workload of 16 VMs.
We noticed that Iterated-KSP is 7.55% better than theIterated-EC (average case) which can be translated into adifference of 1.66 KW.
16 / 24
Energy-Efficient VMs
Placement
Albert De LaFuente
VigliottiDaniel
MacedoBatista
The Problem
The objective
Motivation
Related Work
ThepyCloudSimFramework
Experiments
Results andConclusions
References
Power consumption comparison - Trace 2 (15.9337,34.7465)
Figure : Power consumption [100 hosts / Trace 2]
17 / 24
Energy-Efficient VMs
Placement
Albert De LaFuente
VigliottiDaniel
MacedoBatista
The Problem
The objective
Motivation
Related Work
ThepyCloudSimFramework
Experiments
Results andConclusions
References
Conclusions - Hardware usage
The Iterated-KSP optimizes hardware by 6.20% to20.40% however it is not stable.
Iterated-EC ranges from 7.49% to 13.14% with a trendto be stable ≈11%.
18 / 24
Energy-Efficient VMs
Placement
Albert De LaFuente
VigliottiDaniel
MacedoBatista
The Problem
The objective
Motivation
Related Work
ThepyCloudSimFramework
Experiments
Results andConclusions
References
Suspended physical hosts comparison - Trace 2(15.9337, 34.7465)
Figure : Suspended physical hosts [100 hosts / Trace 2]
19 / 24
Energy-Efficient VMs
Placement
Albert De LaFuente
VigliottiDaniel
MacedoBatista
The Problem
The objective
Motivation
Related Work
ThepyCloudSimFramework
Experiments
Results andConclusions
References
Execution time comparison - Trace 3 (15.1083,44.7083)
Figure : Suspended physical hosts [100 hosts / Trace 3]
20 / 24
Energy-Efficient VMs
Placement
Albert De LaFuente
VigliottiDaniel
MacedoBatista
The Problem
The objective
Motivation
Related Work
ThepyCloudSimFramework
Experiments
Results andConclusions
References
Conclusions - Time
The Iterated-KSP is 11% to 15% faster thanIterated-EC. The execution time difference tend toincrease with the number of hosts and VMs at a rate of≈5 seconds per 100 hosts.
Iterated-EC is easier to be run in parallel thanIterated-KSP
21 / 24
Energy-Efficient VMs
Placement
Albert De LaFuente
VigliottiDaniel
MacedoBatista
The Problem
The objective
Motivation
Related Work
ThepyCloudSimFramework
Experiments
Results andConclusions
References
Thanks
github.com/vonpupp/sbrc-2014-simulation
albert at ime.usp.br
http://www.ime.usp.br/~albert
http://www.albertdelafuente.com
22 / 24
Energy-Efficient VMs
Placement
Albert De LaFuente
VigliottiDaniel
MacedoBatista
The Problem
The objective
Motivation
Related Work
ThepyCloudSimFramework
Experiments
Results andConclusions
References
References I
[1] A. Beloglazov, R. Buyya, Y. C. Lee, and A. Zomaya, “A taxonomy and surveyof energy-efficient data centers and cloud computing systems,” arXiv e-print1007.0066, Jul. 2010. [Online]. Available: http://arxiv.org/abs/1007.0066
[2] Rich Brown, “Report to congress on server and data center energyefficiency:Public law 109-431,” 2007. [Online]. Available:http://www.energystar.gov/ia/partners/prod development/downloads/EPA Datacenter Report Congress Final1.pdf
[3] M. Xavier, M. Neves, F. Rossi, T. Ferreto, T. Lange, and C. De Rose,“Performance evaluation of container-based virtualization for highperformance computing environments,” in 2013 21st Euromicro InternationalConference on Parallel, Distributed and Network-Based Processing (PDP),2013, pp. 233–240.
[4] J. Mehnert-Spahn, E. Feller, and M. Schoettner, “Incremental checkpointingfor grids,” in Linux Symposium, vol. 120, 2009. [Online]. Available:https://www.kernel.org/doc/ols/2009/ols2009-pages-201-208.pdf
23 / 24
Energy-Efficient VMs
Placement
Albert De LaFuente
VigliottiDaniel
MacedoBatista
The Problem
The objective
Motivation
Related Work
ThepyCloudSimFramework
Experiments
Results andConclusions
References
References II
[5] A. Beloglazov and R. Buyya, “OpenStack neat: A framework for dynamicconsolidation of virtual machines in OpenStack clouds–A blueprint,”Technical Report CLOUDS-TR-2012-4, Cloud Computing and DistributedSystems Laboratory, The University of Melbourne, Tech. Rep., 2012. [Online].Available:http://www.cloudbus.org/reports/OpenStack-neat-Blueprint-Aug2012.pdf
[6] R. N. Calheiros, R. Ranjan, A. Beloglazov, C. A. F. De Rose, and R. Buyya,“CloudSim: a toolkit for modeling and simulation of cloud computingenvironments and evaluation of resource provisioning algorithms,” Software:Practice and Experience, vol. 41, no. 1, pp. 23–50, Jan. 2011. [Online].Available: http://onlinelibrary.wiley.com/doi/10.1002/spe.995/abstract
24 / 24